You are on page 1of 23

GEOMETRIC CORRECTION OF LANDSAT MSS, TM, AND ETM+ IMAGERY FOR MAPPING OF WOODY VEGETATION COVER AND CHANGE

DETECTION IN QUEENSLAND
Armston, J.D., Danaher, T.J., Goulevitch, B.M., and Byrne, M.I.

Climate Impacts and Natural Resource Systems Natural Resource Sciences Queensland Department of Natural Resources and Mines 80 Meiers Road Indooroopilly 4068 Queensland Australia Phone: 61-7-38969696, Fax: 61-7-38969606 John.Armston@nrm.qld.gov.au www.nrm.qld.gov.au/slats Abstract The Statewide Landcover and Trees Study (SLATS) requires accurate and precise geometric correction of Landsat Multi-Spectral Scanner (MSS), Thematic Mapper (TM), and Enhanced Thematic Mapper (ETM+) scenes covering Queensland for radiometric calibration, mapping of woody vegetation cover, change detection analysis, and mosaicing of products. In this study, the SLATS geometric correction methodology and results are presented. The use of alternative empirical and photogrammetric models, within the functionality of commercially available image processing packages, is evaluated. Systematic non-linear errors that exhibit significant high order polynomial or high spatial frequency forms which remain in Landsat-7 ETM+, Landsat-5 MSS and Landsat-1 MSS imagery following orthorectification are examined. Emphasis is placed on the spatial assessment of absolute and relative geometric errors. Continuation of this preliminary investigation, intended future research towards improving the SLATS geometric correction of Landsat imagery, and the implications for mapping of woody vegetation cover and change detection are discussed. Introduction The Statewide Landcover and Trees Study (SLATS) was initiated by the Queensland Government in 1995 to provide quantitative data on woody vegetation cover throughout Queensland (Danaher et al., 1998). SLATS provides mapping of woody foliage projective cover (FPC) (Goulevitch et al., 2002), changes in woody vegetation cover (Wedderburn-Bisshop et al., 2002) and woodland thickening (Burrows et al., 2002). These data have been subsequently utilised by various stakeholders for remnant vegetation mapping (Accad et al., 2001), policy formulation and tree management planning, estimation of greenhouse gas emissions (Henry et al., 2002), natural resource assessment and other associated research. Landsat-1, -2, -3, -5 Multi-Spectral Scanner (MSS), Landsat-5 Thematic Mapper (TM), and Landsat-7 Enhanced Thematic Mapper (ETM+) scenes covering the entire state of Queensland have been acquired for sequential time periods from 1972 to 2001. SLATS has developed an integrated Landsat imagery and field measurement based methodology. It relies on quantitative spatial and temporal analysis of these data and includes radiometric calibration (Danaher et al., 2001), change detection analysis, mapping of woody FPC (Goulevitch et al., 2002), and mosaicing of imagery and

ISBN 0-9581366-0-2

11ARSPC

thematic products (Walls et al., 1998). Hence, it is essential that all scenes used in the study are accurately and precisely orthorectified to minimise misregistration among spatially adjacent scenes and between temporal sequences of imagery. Considering the large quantity of Landsat imagery archived by SLATS and its implementation in change detection, a spatially explicit and automated means of assessing systematic geometric errors in the imagery is required. The primary aim of this study is to present the current geometric correction methodology employed by SLATS for the pre-processing of Landsat MSS, TM, and ETM+ imagery, and conduct a preliminary analysis of observed geometric errors. Commonly used approaches to the geometric correction of Landsat imagery, both empirical and physical, are briefly evaluated and compared within the functionality of commercially available image processing packages. The current accuracy of the SLATS geometric correction, the form of systematic non-linear errors remaining in the orthorectified imagery, and the magnitude and direction of image-to-image misregistration are quantified and discussed. Geometric Correction of Landsat Imagery Geometric correction of satellite images involves modelling the relationship between the image and ground coordinate systems. There are both systematic and non-systematic geometric errors present in satellite imagery (Jenson, 1996). The systematic errors in Landsat imagery are well documented, and are primarily functions of scan skew, mirrorscan velocity, panoramic distortion, platform velocity, perspective and earth rotation (Mather, 1999; Jenson, 1996). Data on sensor characteristics and ephemeris information are modelled and applied to the raw imagery as part of the systematic correction performed by the Landsat receiving stations (Masek et al., 2001). Assuming an accurate ephemeris based correction software model is implemented, systematic errors are corrected in commercially available Landsat imagery (e.g. ACRES Landsat-7 ETM+ Level 5 product). Non-systematic errors are mainly caused by variation through time in the position and attitude angles of the satellite platform (Jenson, 1996). Without accurate sensor platform orientation parameters, these errors can only be corrected in image-to-map rectification or image-to-image registration with the use of Ground Control Points (GCPs) and a suitable precision photogrammetric or empirical model (Mather, 1999; Jenson, 1996). Photogrammetric rectification consists of modelling the exterior and interior orientation parameters of the sensor image acquisition geometry in relation to ground coordinates. This requires sensor dependent orbital models which are used to establish a collinear relationship between the sensor-to-image and image-to-ground coordinate vectors (Yang, 1997; Bannari et al., 1995; Pal and Pons, 1995; Toutin, 1995; and Novak, 1992). The condition of collinearity is a fundamental photogrammetric principal and the photogrammetric software models used in this study incorporate collinearity equations. The Landsat sensors have an optical-mechanic line scanner that acquires each pixel individually. Hence, each pixel has a unique value for the exterior orientation parameters of attitude angles (omega, phi, and kappa) and coordinates of the sensor platform (x, y, and z perspective centre), which are all functions of time (Yang, 1997; Bannari et al., 1995). As high precision ephemeris and attitude information is not publicly available for the Landsat sensors, the exterior orientation parameters must be approximated by using Ground Control Points (GCPs) and by means of iterative least

ISBN 0-9581366-0-2

11ARSPC

squares adjustment to achieve precise rectification (Yang, 1997). However, it is important to note that the ephemeris and attitude information for Landsat-7 ETM+ is vastly improved compared to its predecessors. This modelling of satellite position and orientation parameters can be extended to simultaneously integrate further platform parameters (e.g. radial velocity), sensor parameters (e.g. instantaneous field of view, orbital eccentricity), Earth related parameters (e.g. Earth-centre-to-satellite distance), and parameters related to the cartographic map projection (Bannari et al., 1995; Toutin, 1995). Empirical geometric correction models are based on the positional relationship between points on a satellite image and points on a map, or points derived from a GPS. These rectification methods include conventional polynomials and the Rational Function Model (RFM). They are usually applied when there are no ephemeris data or other explicit information on the sources of geometric distortions, as they correct the geometric distortions in an image simultaneously. Furthermore, their mathematical simplicity, computational efficiency, and ability to correct distortions that are not accounted for by sensor orbital models make them an attractive alternative to photogrammetric models. However, as these empirical models only correct locally at GCPs, a high number of GCPs is required to adequately model the distortions in an image, and polynomial approximation away from GCPS will vary more with increasing order of the function (Mather, 1999; Yang, 1997; Bannari et al., 1995; Novak, 1992). First order conventional polynomials allow for a translation, a rotation, and a scaling correction in both axes of an image. Higher order polynomials allow for correction of larger and non-linear distortions (Bannari, 1995; Pal and Pons, 1995; Toutin, 1995). The RFM employs a ratio of two polynomials to approximate image row coordinates, and another to approximate image column coordinates. Each polynomial is a function of normalised point coordinates (x, y, z) (Tao and Hu, 2001). The RFM has advantages over conventional polynomials due to its independence of sensor specific specifications, ability to correct for relief displacement, and smoother interpolation properties between GCPs (Tao and Hu, 2001). To perform change detection analysis, regardless of approach, temporal sequences of images need to be geometrically registered to sub-pixel accuracy to avoid spurious results when detecting changes in reflectance for land cover (Phinn and Rowland, 2001; Wedderburn-Bisshop et al., 2002). Depending on the aim of the change detection, several authors have suggested differing magnitudes of acceptable root mean square (RMS) error. The SLATS change detection methodology specifies a RMS error of less than 0.7 pixels (Wedderburn-Bisshop et al., 2002), and Jenson (1996) specifies an RMS error of less than 0.5 pixels. In a study of the impact of misregistration on change detection, Townshend et al. (1992) found that to avoid significant variance in pixel values a RMS error of less than 0.2 pixels was required for densely vegetated areas, and a RMS error between 0.5 and 1 pixel was necessary for sparsely vegetated areas. Therefore, misregistration is an important factor in change detection studies, so an effective means of evaluation is necessary. The spatial analysis of absolute and relative geometric error has recently been examined by Phinn and Rowland (2001). They introduced a spatially explicit analysis of misregistration to investigate systematic trends in misregistration and consequential spurious land cover change. Consideration and assessment of misregistration are required for consistent and accurate mapping of

ISBN 0-9581366-0-2

11ARSPC

land cover and change detection at all scales of investigation (Phinn and Rowland, 2001; Townshend et al., 1992). Methodology
Data Acquisition

Landsat MSS scenes covering Queensland were acquired for the periods 1972 to 1978, 1979 to 1981, and 1984 to 1985. All Landsat-1 and -2 MSS scenes prior to 1979 were purchased from the United States Geological Survey (USGS) as a Level 8 product. This was the only product format available at the time of purchase. This imagery was acquired in NLAPS format resampled to 57 by 57 metre pixels. Landsat-3 MSS scenes were acquired in CCRS format from ACRES. All post-1979 Landsat-1, -2 and -5 MSS imagery was purchased from the Australian Centre for Remote Sensing (ACRES) as a Level A1 product resampled to 57 by 79 to 82 metre pixels depending on the preprocessing software system. In 1988, ACRES commenced routine acquisition of Landsat-5 TM and its successor Landsat-7 ETM+. Landsat-TM scenes covering Queensland for the periods 1987 to 1988, 1990 to 1991, 1994 to 1995, and 1997 to 1998 were purchased from ACRES as a Level 5 product in CCRS format and resampled to 30 by 30 metre pixels. All Landsat-7 ETM+ scenes covering Queensland for the periods 1999 to 2000 and 2001 were purchased from ACRES as a Level 5 product in FastL7A format. Preprocessing of all Landsat imagery involved systematic georeferencing and the use of a cubic convolution resampling kernel. The Landsat scenes analysed in this study are listed in Table 1.
Table 1. Landsat scenes analysed in this study. Scene centre is in latitude/longitude (decimal degrees). Location Bransby Quilpie Quilpie Quilpie Noccundra Scene Centre -28.39/142.19 -26.94/144.12 -26.94/144.12 -26.00/144.19 -27.36/142.98 Acquisition Date 27/09/2001 16/07/2001 14/08/1999 12/08/1984 20/09/1972 Sensor Landsat-7 ETM+ Landsat-7 ETM+ Landsat-7 ETM+ Landsat-5 MSS Landsat-1 MSS Distributor ACRES ACRES ACRES ACRES USGS

Collection of Ground Control Points

Field collection of SLATS GCPs was conducted throughout Queensland during the period of 1995 to 1997. For coastal scenes that were predominantly ocean, GCPs were also collected from various topographic reef maps. An independent set of GCPs for the Landsat-7 ETM+ Quilpie scene analysed in this study were collected in May 2002. Field collection of GCPs was carried out using a real time Differential Global Positioning System (DGPS) mounted in a 4WD vehicle. The system consisted of a Fugro Omnistar demodulator that received a differential signal from the Optus AUSSAT satellite, and was linked to a Garmin 12XL hand-held GPS fixed inside the vehicle. The DGPS provided a spatial accuracy of less than 10 metres 95% of the time (Murray Bannister, pers.comm.). GCP coordinates were logged until a minimum of 200

ISBN 0-9581366-0-2

11ARSPC

points had been recorded to achieve a confident level of accuracy with minimal standard deviation RMS error of less than 5 metres. Accuracy was dependent on factors such as satellite coverage and geometry, atmospheric conditions, distance from differential base station used and line of sight to the satellites. The real time DGPS navigation system set up in the vehicle was linked to a laptop computer that allowed field data for the GCP to be logged on site using the Custom Information Gathering System (CIGS) software. Information collected and logged on site included the following: a) b) c) d) e) A unique GCP identifier; Site description including vegetation, soils, land use and infrastructure; Compass bearing from the GCP site to the DGPS in degrees magnetic; Distance from the GCP site to the DGPS in metres; Digital photo taken of the GCP site.

Landsat imagery of the field site was displayed on the laptop computer using CIGS, allowing on site assessment of the ability to interpret the position of each GCP on the Landsat image. In addition, a field diagram was produced on site showing the exact GCP position and the geometric configuration of surrounding ground features.
The 1999 Landsat-7 ETM+ Geometric Baseline of Queensland

The creation of the SLATS geometric baseline involved the orthorectification of 1999 Landsat-7 ETM+ imagery covering the entire state of Queensland. The specifications for the creation of the 1999 Landsat-7 ETM+ geometric baseline (henceforth referred to as the geometric baseline), Tie Point (TP) collection and subsequent image-to-image registration closely followed those of the National Carbon Accounting System developed by the Australian Greenhouse Office (Furby, 2001). The SLATS methodology takes advantage of the higher geodetic accuracy of the Landsat-7 ETM+ sensor and higher resolution Landsat-7 ETM+ 15m panchromatic band, as opposed to Landsat-5 TM (Masek et al., 2001). The panchromatic, multispectral, and thermal band pixels of different sizes were co-aligned by assigning nominal pixel sizes of 0.5, 1 and 2, respectively, so that the geometric transformation calculated for the panchromatic band could be applied to all bands. Imagery was corrected for relief displacement using a merged 3 and 9 digital elevation model (DEM) of Queensland. This DEM incorporated an AUSLIG 3 DEM covering Queensland and an AULSIG 9 DEM of Australia, for which a subset was created to give coverage west from longitude 138:00:00 to 136:30:30. Single scene orthorectification of the geometric baseline was performed using the ERDAS Imagine 8.4 Landsat Geometric Model, which is derived using space resection based on collinearity equations. All the Landsat-7 ETM+ imagery was transformed from Level 5 scene coordinates to Map Grid of Australia 1994 (MGA94) coordinates. A standardised processing methodology for the orthorectification was developed by SLATS. The position of GCPs in each image were plotted using a 5/8/2 (mid-infrared, panchromatic and green respectively) band combination false colour display. The criteria for accepting a transformation for any scene included: a) Total RMS error less than 20 metres; b) Uniform geometry of GCPs;

ISBN 0-9581366-0-2

11ARSPC

c) Inclusion of the maximum number of GCPs that met the total RMS error criteria. The panchromatic, multi-spectral, and thermal bands for each scene were resampled to MGA94 coordinates using the transformation model and a cubic convolution kernel. The panchromatic, multi-spectral, and thermal bands were resampled to 12.5 metres, 25 metres, and 50 metres respectively. All subsequent Landsat-7 ETM+ scenes were registered directly to their corresponding geometric baseline scene. However, the Landsat-5 TM and Landsat-1, -2, -3, -4 and -5 MSS sensors have differing ground footprints to the Landsat-7 ETM+ preventing direct scene-to-scene registration. The solution was to create four single band Landsat-7 ETM+ (Band 3: 0.63-0.69m) scene mosaics based on MGA94 coordinates in zones 53, 54, 55, and 56.
Tie Point Capture

The registration of all Landsat images to the geometric baseline required the capture of more than 250 TPs for each 1999 Landsat-7 ETM+ scene. A transformation model for the registration of an uncorrected Landsat-7 ETM+ scene to its corresponding geometric baseline scene was established and TPs were then uniformly captured on the geometric baseline scene using the criteria shown in Table 2. The coordinates of the TPs on the uncorrected image were predicted from the transformation model, and automatically matched between the panchromatic bands of the Landsat-7 ETM+ scenes using image chip correlation matching. TPs with a residual greater than 1 pixel or a correlation coefficient less than 0.8 were discarded, and a random selection of remaining TPs were examined to ensure that the correct features were consistently being matched. The final TPs had uniform geometry across the scene, were pattern matched between scenes, and met the maximum possible number of specified criteria listed in Table 2.
Table 2. Criteria used to select Tie Points. Selected Tie Point Attributes High contrast patterns Features maintain spectral definition over time Geometric configuration in 2 orthogonal directions Bare ground patches amongst dense vegetation Features within remnant vegetation Avoided Tie Point Attributes Evidence of land clearing Water bodies (coast, dams, rivers, etc.) Cloud coverage Feature not defined within a 55 pixel window Features subject to shadowing changes

Registration of Landsat MSS, TM, and ETM+ Scenes to the Geometric Baseline

All Landsat-1, -2, -3, -4 and -5 MSS, Landsat-5 TM, and Landsat-7 ETM+ scenes were individually orthorectified to MGA94 coordinates by registration to the geometric baseline. The ERDAS Imagine 8.4 Landsat-7 ETM+ linear Geometric Model, TPs and the merged AUSLIG 9s and 3s DEM were used in the orthorectification process. Landsat-7 ETM+ scenes were registered to their corresponding geometric baseline scenes, and Landsat-5 TM and Landsat MSS scenes were registered to the Landsat-7 ETM+ Band 3 mosaics as previously described. Testing was performed to discern which Landsat-5 TM and Landsat MSS bands resulted in the highest mean chip correlation. The visible red band provided the most contrast and maintained spectral definition through time for features selected as TPs.

ISBN 0-9581366-0-2

11ARSPC

The differences between the registration of Landsat MSS, TM, and ETM+ are summarised in Table 3.
Table 3. Different criteria used for registering Landsat MSS, TM, and ETM+ scenes to the 1999 Landsat7 ETM+ geometric baseline coverage of Queensland. Parameter Image Chip Correlation Matching: Geometric baseline band Uncorrected scene band -1 to -5 MSS -5 TM -7 ETM+

TM Band 3 (red): 0.63-0.69m MSS band 2 (red):0.6-0.7m TM band 3 (red): 0.63-0.69m

ETM+ Band 8 (pan): 0.52-0.90m ETM+ band 8 (pan): 0.52-0.90m

Criteria for Final Transformation: Remaining Tie Points Tie Point geometry Total RMS error Max. residual error Min. correlation match coefficient

Minimum 50 Uniform with no Tie Point clusters or gaps 0.6 pixels 0.6 pixels (18 m) 1.2 pixels (18 m) (~34.2 m) 1 pixel (~57 m) 1 pixel (30 m) 2 pixels (30 m) 0.7 0.8 0.8

Resampling: Resample kernel Pixel size

50 metres

Cubic Convolution 25 metres Pan: 12.5 m Multi-Spectral: 25 m Thermal: 50 m

An Imagine Landsat Geometric Model was initiated and the coordinates of the previously collected TPs on the uncorrected scene were predicted from the transformation model, and automatically matched to the geometric baseline. TPs not meeting the criteria in Table 3 were discarded and the transformation model was recalculated. This process was iterated until an optimal transformation model was obtained. If there were too few TPs remaining with non-uniform geometry, more were manually captured from appropriate areas of the scene by means of the TP capture methodology previously described. Resampling was performed to project the uncorrected image to MGA94, followed by a qualitative visual validation. This was carried out by overlaying and rapidly alternating the display of the orthorectified multispectral (plus panchromatic and thermal for Landsat-7 ETM+ scenes) bands over their respective geometric baseline bands. A preliminary analysis of variance is conducted for x RMS error, y RMS error, and total RMS error resulting from the orthorectification of Queensland Landsat TM and ETM+ imagery for the periods 1984/1985, 1987/1988, 1990/1991, 1994/1995, 1997/1998 and 2001.
Registration of Landsat MSS Scenes with Systematic Non-linear Errors

Landsat MSS scenes which contain systematic non-linear errors that exhibited significant quadratic and higher order polynomial forms were identified. An interim

ISBN 0-9581366-0-2

11ARSPC

solution based on a polynomial adjustment was applied to the affected scenes prior to orthorectification using the ERDAS Imagine 8.4 Landsat MSS Geometric Models. The image chip correlation matching search radius and search window used were small relative to the range of the systematic non-linear errors in the Landsat MSS scenes. Hence, all the captured TPs needed to be positioned on the uncorrected scenes using a polynomial transformation. The use of a polynomial transformation model, TP prediction, and image chip correlation matching follow the basic methodology described for the registration of Landsat scenes to the geometric baseline. The selection of a polynomial function that best fits the TPs was defined as the polynomial function that provided the greatest proportional decrease in Total RMS error. After the optimal transformation had been calculated, the scene coordinates for each TP were imported into the applicable Imagine Landsat MSS Geometric Model transformation. Once the Landsat MSS orbital model transformation had been calculated, the x and y residuals were plotted against the across and along track raw scene coordinates. Systematic non-linear errors were usually in the along track direction in the x and/or y residuals. These errors were modelled with the order of polynomial that provided the best fit. A polynomial rectification was used to resample the uncorrected Landsat MSS scene to the TP scene coordinates. This removed systematic non-linear errors in the Landsat MSS scenes, enabling accurate and precise orthorectification to be applied by using the ERDAS Imagine 8.4 Landsat MSS linear Geometric Models. Descriptive statistics for x RMS error, y RMS error, and total RMS error resulting from the two orthorectification approaches for the Landsat-1, -2, and -3 MSS scenes are both calculated and compared. This was performed for the period 1979-1981 only, as the 1972-1978 scenes have not been orthorectified to date.
Data Analysis of Geometric Errors

A detailed investigation of geometric errors was carried out for the scenes listed in Table 1. ERDAS Imagine 8.5 Professional and the PCI Geomatica 8.2 OrthoEngine Module, henceforth referred to as Imagine and Geomatica, respectively, were used for all image processing. ERDAS provided a Landsat quadric Geometric Model for testing SLATS Landsat scenes with Imagine 8.5. Quadratic variation in the along track direction for one or more of the attitude angles phi, kappa, and omega in the collinearity equations of the space resection were allowed. Geomatica does not support photogrammetric orthorectification of Landsat MSS imagery. The Quilpie geometric baseline scene, which exhibits minimal relief variation, was identified as a problematic scene in the statewide orthorectification. Hence, a temporal sequence of scenes covering this geographic area was selected for the preliminary data analysis of geometric errors (see Table 1). A comparison of rectification techniques was performed for the 1999 Landsat-7 ETM+ Quilpie scene using 42 SLATS DGPS field surveyed GCPs, and validated using an independent set of 19 DGPS field surveyed Check Points (CPs). The rectification techniques included (a) Imagine Landsat-7 Geometric Models (linear and quadric), (b) Geomatica Landsat-7 orbital model, (c) polynomials (linear, quadratic and cubic), and (d) RFM (5 coefficients and 20 coefficients). The number of coefficients used in each of the two RFMs was determined by: (1) the number of coefficients that provided the

ISBN 0-9581366-0-2

11ARSPC

greatest proportional decrease in total RMS error; and (2) the number of coefficients that provided the lowest total RMS error, respectively. The systematic non-linear errors not modelled by Landsat orbital models in Imagine or Geomatica were quantified. The analysis employed was a polynomial regression with the x or y residuals as the dependent variable, and with along or across track coordinates as the independent variable. Plotting the x or y residuals against the along or across track coordinates enabled systematic non-linear errors, the distribution of residual error, and geometric anomalies (e.g. scene fractures) in the data to be easily identified and visualised. Misregistration between the orthorectified target scenes (see Table 1) and the geometric baseline was quantified. Misregistration between the east-west and north-south overlaps of scenes spatially adjacent to the 2001 Landsat-7 ETM+ Quilpie scene was also quantified. Misregistration was calculated by image chip correlation matching of TPs between scenes. TPs were captured from the geometric baseline as described previously. Easting and northing misregistration RMS error were calculated for each TP using Equation 1. RMS error = (1)

Where xi and yi are the geometric baseline or the target scene easting and northing, and xr and yr are the target scene or target scene overlap easting and northing, respectively. TPs with an image chip correlation coefficient less than 0.8 were discarded and TP outliers were checked to visually interpret if the correct ground feature had been matched. The percentage of TPs that have total RMS errors greater than 1 pixel and 0.5 pixels were also calculated, as these values represented different thresholds that can have a significant effect on automated change detection accuracy (Jenson, 1996; Townshend et al., 1992). The 2001 Landsat-7 ETM+ scene was registered to the geometric baseline scene which was orthorectified to the original SLATS field surveyed DGPS GCPs for Quilpie using the Imagine Landsat quadric orbital model. This model was used as it will be the basis of all future rectifications of SLATS Landsat ETM+ and TM scenes. The Landsat MSS target scenes were registered to the geometric baseline mosaics which were orthorectified using the Imagine Landsat-7 ETM+ linear orbital model. Time constraints prevented the recreation of the 1999 Landsat-7 ETM+ Band 3 mosaics using the Imagine Landsat quadric Geometric Model in the orthorectification process. However, a comparison of the magnitude of geometric error between the Landsat MSS and the geometric baseline suggested that the misregistration results for the Landsat MSS scenes listed in Table 1 would not be confounded by the relatively minor geometric error in the geometric baseline mosaics. Results
Queensland

Global absolute error statistics were calculated for the geometric baseline imagery covering Queensland and are shown in Table 4.

ISBN 0-9581366-0-2

11ARSPC

7DEOH  $EVROXWH HUURU VWDWLVWLFV SHU VFHQH IRU WKH  /DQGVDW (70 JHRPHWULF EDVHOLQH FRYHUDJH RI 4XHHQVODQG RUWKRUHFWLILHG WR UHDO WLPH '*36 PHDVXUHG *&3V XVLQJ WKH ,PDJLQH /DQGVDW (70 OLQHDU RUELWDO PRGHO  VFHQHV IRU WKH JHRPHWULF EDVHOLQH RI 4XHHQVODQG ZHUH SURFHVVHG (UURU XQLWV DUH LQ PHWUHV 4XHHQVODQG ,PDJHU\  /DQGVDW (70 6WDWLVWLF 0HDQ 6W GHY 0D[LPXP 0LQLPXP ;506 HUURU     <506 HUURU     7RWDO 506 HUURU     *&3V    

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
D


E 

506( ( PHWUHV UURU P

< 506( ; 506(



506( PHWUHV ( UURU P



    







   


 70


 70


 70


 70


 (70

066  066  70






 70


 70


 70


 70


 (70

 70 066  066

Era and Sensor of Landsat Imagery

Era and Sensor of Landsat Imagery

ISBN 0-9581366-0-2

10

11ARSPC

Figure 1. Mean (a) Total RMS error (metres), and (b) x and y RMS error (metres), for each Landsat imagery era of Queensland registered to the geometric baseline. The 1975 coverage is excluded due to incomplete rectification. The standard deviation is shown for each era mean. See Appendix Table 1 for numerical results.

A significant difference between the mean RMS error for each Landsat imagery era (F=42.66, p<0.01) is observed in Figure 1a. Pairwise comparisons of 95% confidence intervals between each era reveal that the mean total RMS error of scenes registered to the 1999 Landsat-7 ETM+ geometric baseline of Queensland generally increases with time relative to the era and sensor of the Queensland Landsat imagery. The 2001 Landsat-7 ETM+ imagery RMS error is significantly less than the RMS error for any other era. There is no significant variation between eras of TM imagery RMS error, except for the 1997 era which is significantly less than the 1991 TM era. Figure 1a shows that there is increasing RMS error in the TM eras as the data moves away from the 1999 Landsat-7 ETM+ baseline, suggesting a greater difficulty in image chip correlation matching of ground features. Figure 1b shows that the x RMS error is significantly greater than the y RMS error (t=24.85, p<0.01), and the differences in magnitude follow the same trends as for total RMS error. See Appendix Table 1 for numerical results. Preliminary analysis has indicated that the combined polynomial adjustment and Landsat Geometric Model approach to orthorectification has significantly improved the results for the 1979 to 1981 era of Landsat MSS imagery. The total RMS error of scenes rectified using the combined approach were significantly lower than those scenes rectified using the Landsat Geometric Model only (t=2.79, p<0.01). See Appendix Table 1 for numerical results.
Landsat-7 ETM+ Quilpie Scene

The results of a comparison of different rectification methodologies using the same Landsat-7 ETM+ scene with the same real-time DGPS measured GCPs and CPs is presented in Table 5.
Table 5. Error statistics for rectification of the panchromatic band of the Landsat-7 ETM+ 1999 Quilpie scene to real-time DGPS field surveyed GCPs using various geometric correction techniques. See text for explanation. Validation of the results used an independent set of field surveyed CPs. Error units are in metres. * denotes orthorectification. x RMS Error 16.48 9.38 9.75 14.48 9.57 8.25 GCPs (42) y Total RMS RMS Error Error 6.85 6.38 7.05 7.82 6.58 6.08 17.85 11.34 12.03 19.15 11.61 10.25 Check Points (19) x y Total RMS RMS RMS Error Error Error 19.16 15.08 13.95 20.72 13.92 12.87 11.31 12.23 9.90 11.21 12.00 12.11 22.25 19.42 17.11 23.56 18.38 17.67 All Points (61) x y Total RMS RMS RMS Error Error Error 17.16 10.73 10.80 18.31 10.76 9.21 7.73 7.67 7.80 8.39 7.76 7.50 18.82 13.20 13.32 20.14 13.26 11.88

Rectification Method ERDAS Imagine Geometric Model Linear * Quadric * PCI Geomatica Model * Polynomial Linear Quadratic Cubic Rational Function Model (RFM) * 5 Coefficients 20 Coefficients

9.30 4.50

6.45 3.15

11.32 5.49

13.80 18.75

12.45 21.30

18.59 28.38

10.35 5.85

8.10 5.40

13.14 7.96

ISBN 0-9581366-0-2

11

11ARSPC

The Imagine Landsat-7 Geometric Model, Geomatica Landsat-7 model, and quadratic polynomial all provided comparable results in the RMS statistics and validation using CPs. The Imagine polynomial model does not provide orthorectification, but as the Quilpie scene has minimal relief displacement and there is an adequate number of GCPs, the results are comparable with the orthorectification models. However, a qualitative visual assessment of the 2001 orthorectified scene, has revealed that the global correction of the sensor dependant methods provided better registration of the images in the south-west quadrant of the scene where there were no GCPs or CPs available. Due to the presence of a systematic non-linear error in the x residuals in the along track direction (Figure 1a.), the Imagine Landsat-7 linear model and linear polynomial provided significantly higher RMS error results. This indicates that models that allow for non-linear variation in the along track direction are necessary to remove these distortions. The RFM which uses 5 coefficients provides a comparable result for the GCPs and CPs. However, while the technique enables the correction of relief displacement, it displays similar problems as those for conventional polynomial rectification, in that the interpolation produces spurious results between GCPs. This is clearly shown for the higher order RFM, where the total RMS error is 28.38 metres. All RMS values and their corresponding CP validation values were within 1 pixel (30 metres), however the CP validation values were consistently greater than the GCP values for all rectification models.
(a)
40

(b)

40

X Residual (metres)

20

X Residual (metres)
0 2000 4000 6000 8000 10000 12000

20

-20

-20

-40

-40 0 2000 4000 6000 8000 10000 12000

Scan Line (Along Track) (c) 40

Scan Line (Along Track)


(d)
40

X Residual (metres)

X Residual (metres)
0 2000 4000 6000 8000 10000 12000

20

20

-20

-20

-40

-40 0 2000 4000 6000 8000 10000 12000

Scan Line (Along Track)

Scan Line (Along Track)

ISBN 0-9581366-0-2

12

11ARSPC

Figure 2. Scan line number (along track) against x residual (metres) for the Landsat-7 ETM+ 1999 Quilpie scene rectified to 61 real-time DGPS field surveyed GCPs using: (a) Imagine Landsat-7 model (linear) (r2=0.560, F=36.94, p<0.01); (b) Imagine Landsat-7 Geometric Model (quadric); (c) Geomatica Landsat-7 model; and (d) Imagine quadratic polynomial.

The variation in x residuals across the along track direction is shown in Figure 2a. A systematic quadratic error (r2=0.560, F=36.94, p<0.01) not accounted for by the Imagine Landsat-7 ETM+ linear Geometric Model can be identified. Figures 2b-2d show the comparable results provided by the Imagine Landsat-7 ETM+ quadric orbital model, Geomatica Landsat-7 ETM+ orbital model, and an Imagine quadratic polynomial where the systematic quadratic error was modelled.

(b)
X Residual (metres)

40

20

-20

-40 0 2000 4000 6000 8000 10000 12000

Scan Line (Along Track)

Figure 3. Misregistration between 423 TPs evenly distributed and pattern matched on the geometric baseline and the orthorectified Landsat-7 ETM+ 2001 Quilpie scene. (a) Scatter plot of the spatial distribution of misregistration between scenes (post resampling), and (b) x residuals in the along track direction (cubic polynomial function fitted).

It is seen from Figure 3b, that there exists high order systematic errors in some LandsatETM+ scenes. A high frequency error and a significant cubic polynomial (R2=0.282, F=55.24, p<0.01) trend is observed. The spatial distribution of misregistration (Figure 3a) demonstrates the effect of the above non-linear error in the south-west corner of the scene overlap, for which there was a deficiency of GCPs in the geometric baseline. Despite the global correction of the photogrammetric models, they are still sensitive to the distribution of GCPs. The observed high frequency error has been observed consistently throughout the SLATS Landsat-7 ETM+ imagery, the most pronounced to date being that for Bransby as shown in Figure 4.

ISBN 0-9581366-0-2

13

11ARSPC

40

X Residuals (metres)

20

-20

-40 0 2000 4000 6000 8000 10000 12000

Scan Line (Along Track)

Figure 4. x residuals in the along track direction for the 2001 Landsat-7 ETM+ Bransby scene registered to the geometric base using TP and Imagine Landsat-7 ETM+ quadric Geometric Model. The high frequency error is evident.

The magnitude of misregistration in east-west and north-south scene overlaps, between adjacent scenes in the easting and northing directions, are shown in Figure 5. There is a larger range of misregistration values in the easting direction for both the north-south and east-west overlaps, however, it is much more profound in the east-west overlap (Figure 5a). There is a substantial number of local misregistration measurements greater than 25 metres (1 pixel) in the east-west overlap.
(a) 40
30

(b) 40
30

Northing Difference (m)

10 0 -10 -20 -30 -40 -40 -30 -20 -10 0 10 20 30 40

Northing Difference (m)

20

20 10 0 -10 -20 -30 -40 -40 -30 -20 -10 0 10 20 30 40

Easting Difference (m)

Easting Difference (m)

Figure 5. Scatter plots of misregistration in the easting and northing directions for the orthorectified 1999 Landsat-7 ETM+ Quilpie scene: (a) east-west overlap with the 1999 Landsat-7 ETM+ Charleville scene; and (b) north to south overlap with the 1999 Landsat-7 ETM+ Adavale scene.

ISBN 0-9581366-0-2

14

11ARSPC

Table 6. Misregistration statistics for the 2001 Landsat-7 ETM+ Quilpie scene. This scene was registered to the geometric baseline of Queensland using the Imagine Landsat-7 ETM+ Geometric Model (quadric) and 423 scene-to-scene correlation matched TPs. Error units are in metres (1 pixel = 25 metres). RMS error Statistic Mean Standard Deviation Maximum Minimum % RMS Error > 1 Pixel % RMS Error > 0.5 Pixel Result 11.93 8.57 57.70 0.52 7.33 36.64

The misregistration statistics between the 1999 and 2001 target Quilpie scenes are summarised in Table 6. It shows that while the mean is less than 0.5 pixels (15 metres), there are locations where the misregistration exceeds 2 pixels (maximum=57.70 metres), thus potentially these have a significant effect on scene-to-scene change analysis. Less than 10% of the local misregistration measurements have magnitudes greater than 1 pixel, however, at least 36% have a magnitude greater than 0.5 pixels.
Landsat-5 MSS Quilpie Scene

The Landsat-5 MSS imagery acquired from ACRES have exhibite numerous anomalies, the most consequential for geometric correction being a 2 or 4 pixel fracture in the centre of the scene. Whilst this is easily removed by subsetting and offsetting image segments, it is also easily overlooked, especially when evaluating the accuracy of the geometric correction with global RMS error statistics. Forty-three Landsat-5 MSS scenes acquired from ACRES prior to 2000 have been identified as having a 2 or 4 pixel line fracture in the centre of the scene. Some Landsat-5 MSS scenes acquired from ACRES also exhibit systematic non-linear errors in the across track direction as illustrated by Figure 6.
(a)
100

(b)

100

X Residuals (metres)

X Residual (metres)

50

50

-50

-50

-100 0 500 1000 1500 2000 2500 3000 3500

-100 0 500 1000 1500 2000 2500 3000 3500

Pixel Line (Across Track)

Pixel Line (Across Track)

Figure 6. Scatter plot of the Landsat-5 MSS Quilpie scene showing x residuals in the across track direction (x) resulting from the registration of the scene to the geometric baseline of Queensland using pattern matched TPs and the quadric Landsat-5 MSS Geometric Model in Imagine 8.5. (a) Imagine Landsat-5 MSS quadric Geometric Model, and (b) RFM using 13 coefficients.

ISBN 0-9581366-0-2

15

11ARSPC

The non-linear trend in the x residuals in the across track direction for (Figure 7a) the Imagine Landsat-5 MSS quadric model is significant (r2 = 0.558, F= 54.071, p<0.05). As the rigidity of available Landsat orbital models does not allow for correction of this non-linear error, the RFM was applied to the scene (Figure 7b). A RFM using 13 coefficients provided the greatest proportional decrease in total RMS error (x RMS=21.09, y RMS=16.80, and total RMS error=26.96 metres). Whilst the RMS error statistics have been reduced, there is significant noise in the data, and the non-linear error is not modelled by the RFM. Furthermore, some TPs have been removed from the RFM due to outlying approximations, and there are still spurious approximations at the edges of the image (scan lines 1 to 500).
Table 7. Misregistration statistics for the 1984 Landsat-5 MSS Quilpie scene. This scene was registered to the geometric baseline of Queensland using: (a) the Imagine Landsat-7 ETM+ Geometric Model (quadric); and (b) rational functions with 13 coefficients, with 356 scene-to-scene correlation matched TPs. Error units are in metres (1 pixel = 50 metres). RMS error Statistic Mean Standard Deviation Maximum Minimum % RMS Error > 1 Pixel % RMS Error > 0.5 Pixel Orthorectification Method Orbital Model RFM 26.69 53.55 14.25 15.48 71.24 96.62 1.64 8.09 7.87 59.04 49.16 96.89

The misregistration statistics, after resampling between the 1984 Landsat-5 MSS Quilpie scene and the geometric baseline using two rectification models, are shown in Table 7. As shown in Figure 3, neither method modelled the systematic non-linear error in the x residuals in the across track direction. However, based on the mean misregistration, the Imagine Landsat-5 MSS quadric Geometric Model provides a significantly better result. This may be due to the high oscillation of the RFM approximations away from the TPs. The RFM result is clearly unacceptable for the SLATS orthorectification of Landsat-5 MSS imagery as over 50% of local misregistration measurements are greater than 1 pixel, as opposed to less than 10% misregistration provided by the Imagine Landsat-5 MSS quadric Geometric Model.
Landsat-1 MSS Noccundra Scene

The plots in Figure 8 show that there are systematic errors exhibiting a cubic form in the 1972 Landsat-1 MSS Noccundra scene. The trend in the y residuals in the along track direction (Figure 8a) is significant (r2=0.690, F=196.65, p<0.01), as is the trend in the x residuals in the along track direction (Figure 8b), despite a large amount of noise and hence a low regression coefficient (r2=0.094, F=9.15, p<0.01). These systematic nonlinear errors exhibit quadratic and cubic forms with varying magnitudes in all the Landsat-1, -2, and -3 imagery acquired from ACRES and USGS.

ISBN 0-9581366-0-2

16

11ARSPC

(a)

150 100

(b)

150 100

Y Residual (metres)

X Residual (metres)

50 0 -50 -100 -150 0 500 1000 1500 2000 2500 3000 3500

50 0 -50 -100 -150 0 500 1000 1500 2000 2500 3000 3500

Scan Line (Along Track)


(c)
150 100

Scan Line (Along Track)

(d)

150 100

Y Residual (metres)

50 0 -50 -100 -150 0 500 1000 1500 2000 2500 3000 3500

X Residual (metres)

50 0 -50 -100 -150 0 500 1000 1500 2000 2500 3000 3500

Scan Line (Along Track) (e)


150 100

Scan Line (Along Track)


(f)
150 100

Y Residual (metres)

X Residual (metres)

50 0 -50 -100 -150 0 500 1000 1500 2000 2500 3000 3500

50 0 -50 -100 -150 0 500 1000 1500 2000 2500 3000 3500

Scan Line (Along Track)

Scan Line (Along Track)

Figure 8. Scatter plots of 1972 Landsat-1 MSS Noccundra scene showing: (a) y residuals in the along track direction (y); and (b) x residuals in the along track direction for the quadric Landsat-1 Geometric Model. Results for the RFM are shown in: (c) y residuals in the along track direction (y); and (d) x residuals in the along track direction. Results for the combined polynomial adjustment and linear Imagine Geometric Model approach are shown in: (e) y residuals in the along track direction (y); and (f) x

ISBN 0-9581366-0-2

17

11ARSPC

residuals in the along track direction. Scenes were registered to the geometric baseline of Queensland using pattern matched TPs. Cubic polynomial functions have been fitted.

The 1972 Landsat-1 MSS Noccundra scene orthorectified with the quadric Landsat-1 MSS Geometric Model had a large RMS error of 0.87 (49.59 metres). The combined polynomial adjustment and linear Imagine Geometric Model approach to orthorectification of the Noccundra scene resulted in an RMS error of 0.58 (33.06 metres) and the RFM orthorectification resulted in an RMS error of 0.52 (29.64 metres). The results in Figure 8 indicate that models allowing cubic variation in the along track direction for the Noccundra Landsat-1 MSS scene result in improved registration to the geometric baseline. The RFM modelled the systematic cubic error in the y residuals in the along track direction as there is no significant cubic trend (p>0.05). However, the systematic cubic trend in the x residuals in the along track direction is significant following orthorectification as shown in Figure 8d (r2=0.202, F=21.92, p<0.01). The combined polynomial adjustment and Imagine linear Geometric Model approach resulted in the modelling of the systematic cubic trends, in both the x and y residuals in the along track direction (see Figure 8e,f). Neither had a significant cubic trend (p>0.05). Discussion
Geometric Correction of Landsat Imagery

The SLATS geometric correction methodology currently provides adequate results for the requirements of the woody vegetation mapping and change detection methodologies. However, the results demonstrate that there is considerable scope for improvement as consistent sub-pixel registration of all scenes has not yet been achieved. The results have demonstrated that linearity in the along track direction for sensor perspective centre and attitude angles cannot be assumed. The combined polynomial adjustment/Landsat orbital model and the RFM, seem viable alternatives for precision orthorectification of Landsat MSS imagery that exhibit severe systematic non-linear distortions. However, with a high number of coefficients, RFMs approximations have been observed to vary away from the GCPs/TPs. This was also found by Tao and Hu (2001), who state that the high order RFMs over parameterize and cause instability in the least squares adjustment. The combined polynomial adjustment/Landsat Geometric Model approach is currently the pragmatic solution for the SLATS geometric correction of Landsat MSS imagery. Some high frequency spectral information is lost from resampling the imagery twice, however, the benefits of this approach for change detection outweigh this. As indicated by Yang (1997), one possible solution to the double resampling problem that requires investigation, is to initially store the transformation results of the polynomial adjustment with the image, then recompute the scene coordinates on-the-fly during the Landsat orbital model orthorectification process. The comparison between the commonly used rectification models reveal that the sensor dependent photogrammetric models generally produce higher accuracy transformations, which is consistent with previous work by various authors (e.g. Yang, 1997; Bannari, 1995; Pal and Pons, 1995; Toutin, 1995; Novak, 1992). Assuming a high number of uniformly distributed GCPs or TPs, stable acquisition geometry, and minimal relief displacement over the image, polynomial rectification can produce comparable rectification results (Bannari et al., 1995).

ISBN 0-9581366-0-2

18

11ARSPC

This study highlights systematic non-linear and high frequency errors which appear to be a function of time, and of which the direct cause at this stage is unknown. Furby (2001) also identified high frequency geometric errors and scene fractures in Landsat imagery covering Australia. Some possibilities include geometric distortions caused by variations in satellite position, and attitude not adequately modelled, as is indicated by the high RMS error of models that assume linearity (see Table 5). Propagation of geometric error in the systematic geometric correction preprocessing by ACRES and USGS (e.g. scan gap correction), error introduced in processing with commercially available image processing packages, and operator error in the measurement and plotting of GCPs and TPs are other possible contributors to the observed geometric error. There are a multitude of possible sources of geometric error that are very difficult to systematically investigate without accurate and precise satellite ephemeris and image preprocessing information. The importance of assessing the accuracy of geometric correction with independent CPs and a spatially explicit assessment of geometric error is emphasised. Interpreting the accuracy of a rectification, especially when using an empirical model, based solely on the RMS error of the GCPs can be misleading. Whilst it is impractical for most applications due to costs, the use of CPs to validate the geometric correction model used is necessary (Yang, 1997; Kardoulas et al., 1996). The accuracy of the SLATS geometric correction methodology for scene-to-scene registration is currently assessed qualitatively, which enables any misregistration to be visually identified. Furthermore, it is demonstrated that the global RMS error statistics are of little use in identifying systematic trends in misregistration. RMS error statistics are an aspatial accuracy assessment, hence a spatially explicit analysis of misregistration is needed to augment the RMS error statistics (Phinn and Rowland, 2001). Scatter plots of x and y residuals in the across track and along track directions, or a method similar to that described by Phinn and Rowland (2001), are useful in identifying systematic misregistration. Global accuracy statistics reveal improved geometric correction results for the Landsat7 ETM+ imagery in comparison with Landsat TM imagery (Figure 1), which may be attributed to the increased geometric fidelity of the landsat-7 ETM+ sensor and use of the panchromatic band (Masek et al., 2001). However, geometric error propagation, between overlapping adjacent scenes and between temporal sequences of imagery, is still a significant problem. This error propagation may be minimised is several ways. For example, Toutin et al. (2001) showed the potential of spatio-triangulation using a block bundle adjustment to simultaneously rectify blocks of Landsat-7 ETM+ imagery. This is considered unsuitable or unnecessary as spatially adjacent SLATS scenes are rarely acquired from the same path and date. SLATS acquired GCPs in scene overlaps for the current methodology and ERDAS Imagine does not provide spatio-triangulation for Landsat imagery. However, further testing is necessary to evaluate the applicability of this method to the SLATS image archive. Additionally, uniform geometry of GCPs has proven a critical factor in the accurate rectification of imagery (Figure 3), even where a global photogrammetric rectification model has been used. In the creation of the geometric baseline, TPs could be used to supplement the DGPS field measured GCPs to create a uniform distribution of ground control for each scene.

ISBN 0-9581366-0-2

19

11ARSPC

Implications for Woody Vegetation Change Detection

It is evident from the results that significant misregistration exists between temporal sequences of images over the same location, and between overlapping adjacent scenes. Local misregistration measures among scenes revealed consistently greater than 1 pixel error, even when the absolute error of the geometric baseline was sub-pixel. This potentially has a significant effect on change detection accuracy regardless of the type of change being investigated (Phinn and Rowland, 2001; Jenson, 1996; Townshend et al., 1992). Whilst the analysis of misregistration in this study occurred at the scale of a single scene, the misregistration would likely propagate significant spurious automated change detection results at the scale of Queensland. We can conclude that the raster filter of classified change in woody vegetation according to Wedderburn-Bisshop et al. (2002) is necessary. Greater than 1 pixel misregistration can introduce spurious change in automated change detection, especially in linear features such as roads. It can be argued that misregistration effects on broad scale detection of woody vegetation clearing and regrowth will balance out with errors of omission and commission. However, as Townshend (1992) points out, increasing misregistration between temporal sequences of images may cause greater variability around the mean and greater overestimates of temporal changes in clearing and regrowth.
Future Research

There are currently several avenues of future research directed at improving the SLATS geometric correction methodology. The feasibility of generating a high resolution Queensland DEM using ASTER nadir and backward imagery, and stereographic techniques is currently being investigated. This may enable vastly improved correction of relief displacement, and also allow topographic correction in the radiometric calibration process. For example, Itten and Meyer (1993) demonstrated that matching resolutions of Landsat imagery and a corresponding DEM is an important consideration in geometric correction of Landsat TM imagery. ASTER or SPOT imagery is also attractive for generating a high resolution, geometrically stable, geometric baseline. As the ASTER (TERRA) and SPOT linear array sensors acquire each scan line simultaneously, any distortions associated with optical-mechanical scanner machinery are avoided. Assuming that a geometric baseline of Queensland is relatively error free, inheritance of geometric errors in registered imagery will be minimised.
Conclusion

SLATS has benefited from a large body of research on geometric correction of Landsat imagery, and has subsequently developed a standardised and reasonably accurate methodology for the geometric correction of the SLATS Landsat ETM+, TM, and MSS image archive of Queensland. The study presents a preliminary investigation only. Further research and development is required to improve the accuracy of the geometric correction and to identify and model systematic errors in the imagery. The SLATS methodology for mapping of woody vegetation and change detection in Queensland has allowed for the effects of misregistration, and hence produces results of high geometric accuracy.

ISBN 0-9581366-0-2

20

11ARSPC

Acknowledgments All members of SLATS and CINRS, past and present, for their contribution to the development and application of the SLATS geometric correction methodology, to the data analysis and fieldwork for this study, and for comments on the manuscript. References Accad, A., Neldner, V.J., Wilson, B.A., and Niehus, R.E. 2001. Remnant Vegetation in Queensland: Analysis of Pre-clearing, Remnant 1997-1999 Regional Ecosystem Information. Queensland Herbarium, Environmental Protection Agency, Brisbane. Bannari, A., Morin, D., Benie, G.B. and Bonn, F.J. 1995. A theoretical review of different mathematical models of geometric corrections applied to remote sensing images. Remote Sensing Reviews, 13: 27-47. Burrows, W.H., Henry, B.K., Back, P.V., Hoffmann, M.B., Tait, L.J., Anderson, E.R., Menke, N., Danaher, T., Carter, J.O., and McKeon, G.M. 2002. Growth and carbon stock change in eucalypt woodlands in northeast Australia: ecological and greenhouse sink implications. Global Change Biology, 8: 769-784. Danaher, T.J., Wedderburn-Bisshop, G.R., Kastanis, L.E., and Carter, J.O. 1998. The Statewide Landcover and Trees Study (SLATS) monitoring land cover change and greenhouse gas emissions in Queensland. Proceedings of the 9th Australasian Remote Sensing and Photogrammetry Conference, Sydney, Australia, July 1998. Danaher, T., Xiaoliang, W., and Campbell, N. 2001. Bi-directional reflectance distribution function approaches to radiometric calibration of Landsat ETM+ imagery. IGARSS 01, Proceedings of the 2001 IEEE International Geoscience and Remote Sensing Symposium, July 2001, Sydney, Australia. Furby, S. 2001. Land Cover Change: Specifications for Remote Sensing Analysis. National Carbon Accounting System Technical Report No. 9 Pre-publication draft, Australian Greenhouse Office, Canberra. Goulevitch, B.M., Danaher, T.J, Stewart, A.J., Harris, D.P., and Lawrence, L.J. 2002. Mapping woody vegetation cover over the state of Queensland using Landsat TM and ETM+ imagery. Submitted to: Proceedings of the 11th Australasian Remote Sensing and Photogrammetry Conference, Brisbane, Australia, September 2002. Henry, B.K., Danaher, T.J., McKeon, G.M., and Burrows, W.H. 2002. A review of the potential role of greenhouse gas abatement in native vegetation management in Queenslands rangelands, Rangelands Journal, 24(1): 112-132. Itten, K.I. and Meyer, P. 1993. Geometric and radiometric correction of TM data of mountainous forested areas. IEEE Transactions on Geoscience and Remote Sensing, 31(4): 764-770. Jenson, J.R. 1996. Introductory Digital Image Processing: A Remote Sensing Perspective, 2nd Edition, Prentice Hall, New Jersey. Kardoulas, N.G., Bird, A.C., and Lawan, A.I. 1996. Geometric correction of SPOT and Landsat imagery: a comparison of map- and GPS-derived control points. Photogrammetric Engineering and Remote Sensing, 62(10): 1173-1177.

ISBN 0-9581366-0-2

21

11ARSPC

Masek, J.G., Honzak, M., Goward, S.N., Liu, P., and Pak, E. 2001. Landsat-7 ETM+ as an observatory for land cover: initial radiometric and geometric comparisons with Landsat-5 Thematic Mapper. Remote Sensing of Environment, 78: 118-130. Mather, P.M. 1999. Computer Processing of Remotely Sensed Images: An Introduction, 2nd Edition, John Wiley & Sons, West Sussex. Novak, K. 1992. Rectification of digital imagery. Photogrammetric Engineering & Remote Sensing, 58(3): 339-344. Pal, V., and Pons, X. 1995. Incorporation of relief in polynomial-based geometric corrections. Photogrammetric Engineering & Remote Sensing, 61(7): 935-944. Phinn, S.R. and Rowland, T. 2001. Quantifying and visualising geometric misregistration from Landsat Thematic Mapper imagery and its effects on change-detection in a rapidly urbanising catchment. Asian-Pacific Remote Sensing & GIS Journal. In Press. Toa, C.V. and Hu, Y. 2001. A comprehensive study of the Rational Function Model for photogrammetric processing. Photogrammetric Engineering & Remote Sensing, 67(12): 1347-1357. Toutin, T. 1995. Multisource data integration: comparison of geometric and radiometric methods. International Journal of Remote Sensing, 16(15): 2795-2811. Toutin, T., Carbonneau, Y., and Chenier, R. 2001. Block adjustment of Landsat-7 ETM+ images. ISPRS Joint Workshop High Resolution from Space, Hanover, Germany, September 19-21, 2001, CD-ROM. Townshend, J.R.G., Justice, C.O., Gurney, C., and McManus, J. 1992. The impact of misregistration on change detection. IEEE Transactions on Geoscience and Remote Sensing, 30(5): 1054-1060. Wedderburn-Bisshop, G., Walls, J., Senerath, U.G. and Stewart, A.J. 2002. A methodology for mapping change in woody landcover in Queensland from 1999 to 2001 using Landsat ETM+. Submitted to: Proceedings of the 11th Australasian Remote Sensing and Photogrammetry Conference, Brisbane, Australia, September 2002. Walls, J., Kastanis, L., and Plunkett, J. 1998. Development of integrated products for tree management planning. Proceedings of the 9th Australasian Remote Sensing and Photogrammetry Conference, Sydney, Australia, July 1998. Yang, X. 1997. Georeferencing CAMS Data: Polynomial Rectification and Beyond, Dissertation, University of South Carolina, Columbia.

ISBN 0-9581366-0-2

22

11ARSPC

Appendix
Table 1. Absolute error statistics per scene for the 1980, 1985, 1988, 1991, 1995, 1997, and 2001 Landsat-7 ETM+, Landsat-5 TM and Landsat MSS era of imagery over Queensland. All scenes were registered to the geometric baseline coverage of Queensland using the Imagine Landsat linear Geometric Models and pattern matched TPs. 87 scenes for each of the Landsat-7 ETM+, 88 scenes for each of the Landsat-5 TM coverages were processed, and 81 out of 91 scenes for the 1985 Landsat-5 MSS coverage. and 76 out of 100 scenes for the 1980 Landsat-2 and -3 MSS coverage. The 1979 to 1981 Landsat-1, -2, and -3 MSS scenes registered to the 1999 Landsat-7 ETM+ geometric baseline coverage of Queensland using (i) the Imagine Landsat MSS linear Geometric Models, and (ii) a combined polynomial adjustment and Landsat MSS linear Geometric Model rectification approach. 100 Landsat scenes constitute the Landsat-2 and -3 MSS scenes for the 1979 to 1981 era. The number of Landsat MSS scenes completed for each rectification method to date are shown. Error units are in metres. Queensland Coverage 2001 Landsat-7 ETM+ Statistic Mean St. Dev. Maximum Minimum Mean St. Dev. Maximum Minimum Mean St. Dev. Maximum Minimum Mean St. Dev. Maximum Minimum Mean St. Dev. Maximum Minimum Mean St. Dev. Maximum Minimum Mean St. Dev. Maximum Minimum Mean St. Dev. Maximum Minimum x Error 12.13 2.36 17.34 7.52 12.78 2.09 19.66 7.03 13.21 1.44 17.92 8.06 13.60 1.44 19.24 10.12 13.57 1.33 18.27 10.12 26.53 2.70 40.36 22.33 28.21 5.68 46.96 19.77 24.87 4.59 32.23 19.82 y Error 7.69 1.61 15.12 4.67 9.05 1.27 12.88 6.37 9.64 1.14 12.49 7.03 9.93 0.93 12.97 8.11 9.69 0.98 12.38 8.11 25.85 3.93 42.85 18.41 36.15 15.56 105.30 20.96 26.95 7.50 36.18 19.89 RMS Error 14.46 2.26 20.32 9.69 15.72 2.01 22.39 9.49 16.40 1.35 20.37 12.07 16.88 1.33 22.21 13.38 16.76 1.19 21.23 13.38 37.17 3.551 51.53 31.60 46.53 14.53 110.48 30.75 36.78 8.24 46.75 28.49 # GCPs 263.57 112.19 493 24 282.49 1.4.76 471 33 286.77 106.23 506 51 269.01 94.84 470 51 262.32 103.82 468 51 240.63 97.99 483 39 188.01 81.87 379 10 194.63 122.16 366 15

1997 Landsat-5 TM

1995 Landsat-5 TM

1991 Landsat-5 TM

1988 Landsat-5 TM

1985 Landsat-5 MSS

1980 Landsat-2,-3 MSS (i)Linear Geometric Model (76 scenes completed)

(ii)Polynomial Adjustment + Linear Geometric Model (8 scenes completed)

ISBN 0-9581366-0-2

23

11ARSPC

You might also like